University of Arizona Health Sciences researchers are harnessing AI and an Oura smart ring to forecast when pregnant women will go into labour.

“Most of us wear health and fitness trackers on our wrists or fingers, and that’s an amazing window into our biology and how we operate,” said Shravan Aras, assistant director at the University of Arizona.

Aras, who is also an assistant research professor, wants to help his fellow scientists get the most out of wearable sensors by incorporating them into research studies and optimising the way wearable data are explored and analysed. The growth of AI and machine learning has opportunities for advancement in this field.

“It’s always been fascinating to me that you can build something out of nothing, like writing code to build tools,” said Aras. “I love being able to write code, do analysis, create the hardware and see it being used by the end users. I’ve always looked at computer science not in isolation, but as a collaborative tool; computer science being applied to different domains to solve really complicated and challenging problems.”

One such problem is accurately predicting when a pregnant woman will go into labour. Due dates are calculated by counting 40 weeks from a woman’s last menstrual period, though in humans, gestation lengths can vary from 37 to 42 weeks. No clinical tools can provide an accurate indication of impending labour, leaving pregnant women to self-report signs of labour, a method with a high rate of false positives. 

When labour appears without warning, maternal health can suffer due to unplanned home births, inadequate time for health care professionals to intervene in preterm births, or recommendations for earlier labour induction if a woman lives far from a hospital, for example.

Certified nurse-midwife Elise Erickson, an associate professor of physiology at the university’s College of Medicine (medicine.arizona.edu), was looking for answers. She invited Aras to collaborate on a study investigating the feasibility of predicting labour onset using temperature data, which is commonly done in animals. In humans, temperature can help determine ovulation and fertility timing.

“Companies that do fertility and ovulation tracking take temperature readings once a day,” Aras said. “There is a relationship between the hormone progesterone and temperature, which is how they can figure out when a person is ovulating. With pregnancy, there are a whole lot of different things going on in the body. It’s not as simple as figuring out if the temperature is going lower or higher. For labour prediction, daily temperature readings do not give you a cohesive pattern of when somebody might go into labour.”

The research team used an Oura ring (ouraring.com) to track temperature readings every minute rather than every day. With an enormous amount of data at his disposal, Aras led efforts to develop a deep neural network-based AI model to analyse the data.

Deep neural networks simulate the activity of the human brain. They are built in layers, with one input layer to accept data and one output layer that generates the result. In between are multiple levels of hidden layers that perform complex calculations, similar to how the brain processes information. In addition to processing the data, deep neural networks learn from the data by using algorithms to compare its prediction with the output and improve accuracy.

By applying deep learning techniques to continuous body temperature data, the researchers were able to accurately predict the day of labour onset. The final model correctly predicted labour start for 79% of spontaneous labours within a 4.6-day window at seven days before true labour, and 7.4-day window at ten days before true labour.

“We were able to develop deep neural network-based AI models that took all of this very high frequency temperature data – one data point per minute of temperature – and come up with a predictive output of when a mother might go into labour,” Aras said.

The team hopes to test the model in a larger study to further refine its clinical applicability. Their goal is to develop software that could be added to existing wearable products or medical devices.

Details of the study can be found at bmcpregnancychildbirth.biomedcentral.com/articles/10.1186/s12884-024-06862-9.